trending/miles
1
0
Fork 0
mirror of https://github.com/radixark/miles.git synced 2026-10-02 07:14:53 +08:00
Code Issues Packages Projects Releases Wiki Activity
Files
3d3fb97bacfac5cc3776ee80382ca30cbca24f7b
miles/tests/snapshots/charts/miles-run/typical-values.yaml
T

333 lines
9.6 KiB
YAML
Raw Blame History

run:
colocate:
inference_pools:
- layout:
gpu_offset: 0
num_gpus_per_inference_pod: 8
num_gpus_per_node: 8
num_inference_cells: 1
num_pods_per_inference_cell: 2
num_pods_per_trainer_cell: 4
num_trainer_cells: 1
pool_id: inference-engine-all-0-0
- layout:
gpu_offset: 16
num_gpus_per_inference_pod: 8
num_gpus_per_node: 8
num_inference_cells: 2
num_pods_per_inference_cell: 1
num_pods_per_trainer_cell: 4
num_trainer_cells: 1
pool_id: inference-engine-all-0-1
namespace: rl
release: myrun
trainer_pool_id: trainer-engine-actor
env:
CUDA_DEVICE_MAX_CONNECTIONS: '1'
PYTHONUNBUFFERED: '1'
id: 260101-000000-000
inferenceEngines:
- command:
- <PYTHON>
- -m
- sglang.launch_server
- --trust-remote-code
- --model-path
- <FIXTURE_DIR>/typical-model
- --host
- 0.0.0.0
- --port
- '8000'
- --device
- cuda
- --random-seed
- '<RANDOM_SEED>'
- --enable-memory-saver
- --nccl-port
- '10000'
- --nnodes
- '2'
- --node-rank
- $(LWS_WORKER_INDEX)
- --dist-init-addr
- $(LWS_LEADER_ADDRESS):9000
- --gated-launch-port
- '13000'
- --tp-size
- '16'
- --skip-server-warmup
- --enable-draft-weights-cpu-backup
- --enable-metrics
- --disaggregation-mode
- prefill
- --load-balance-method
- round_robin
- --disaggregation-bootstrap-port
- '11000'
- --engine-info-bootstrap-port
- '12000'
- --mem-fraction-static
- '0.7'
- --chunked-prefill-size
- '4096'
- --cuda-graph-backend-prefill
- disabled
- --lora-use-virtual-experts
env:
NVSHMEM_DISABLE_NCCL: '1'
RAY_EXPERIMENTAL_NOSET_ASCEND_RT_VISIBLE_DEVICES: '1'
RAY_EXPERIMENTAL_NOSET_CUDA_VISIBLE_DEVICES: '1'
RAY_EXPERIMENTAL_NOSET_HABANA_VISIBLE_MODULES: '1'
RAY_EXPERIMENTAL_NOSET_HIP_VISIBLE_DEVICES: '1'
RAY_EXPERIMENTAL_NOSET_NEURON_RT_VISIBLE_CORES: '1'
RAY_EXPERIMENTAL_NOSET_ONEAPI_DEVICE_SELECTOR: '1'
RAY_EXPERIMENTAL_NOSET_TPU_VISIBLE_CHIPS: '1'
SGLANG_BATCH_INVARIANT_OPS_ENABLE_MM_FALLBACK_VARIANT: 'true'
SGLANG_DG_CACHE_DIR_PER_PROCESS: '1'
SGLANG_ENABLE_HEALTH_ENDPOINT_GENERATION: 'false'
SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_IDLE: 'false'
SGLANG_ENABLE_TP_MEMORY_INBALANCE_CHECK: 'false'
SGLANG_EXPOSE_OWN_ENV_VARS: '1'
SGLANG_JIT_DEEPGEMM_PRECOMPILE: 'false'
SGLANG_MEMORY_SAVER_CUDA_GRAPH: 'true'
SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2: '1'
meta:
gpu_ids: 0,1,2,3,4,5,6,7
name: inference-engine-all-0-0
objectName: myrun-miles-run-inference-engine-all-0-0
poolId: inference-engine-all-0-0
ports:
- name: primary
port: 8000
- name: dist-init
port: 9000
- name: nccl
port: 10000
- name: disaggregation
port: 11000
- name: engine-info-boo
port: 12000
- name: gate
port: 13000
replicas: 1
resources:
limits:
nvidia.com/gpu: 8
size: 2
- command:
- <PYTHON>
- -m
- sglang.launch_server
- --trust-remote-code
- --model-path
- <FIXTURE_DIR>/typical-model
- --host
- 0.0.0.0
- --port
- '8000'
- --device
- cuda
- --random-seed
- '<RANDOM_SEED>'
- --enable-memory-saver
- --nccl-port
- '10000'
- --node-rank
- $(LWS_WORKER_INDEX)
- --dist-init-addr
- $(LWS_LEADER_ADDRESS):9000
- --gated-launch-port
- '13000'
- --tp-size
- '8'
- --skip-server-warmup
- --enable-draft-weights-cpu-backup
- --enable-metrics
- --disaggregation-mode
- decode
- --engine-info-bootstrap-port
- '12000'
- --mem-fraction-static
- '0.7'
- --chunked-prefill-size
- '4096'
- --cuda-graph-backend-prefill
- disabled
- --lora-use-virtual-experts
env:
NVSHMEM_DISABLE_NCCL: '1'
RAY_EXPERIMENTAL_NOSET_ASCEND_RT_VISIBLE_DEVICES: '1'
RAY_EXPERIMENTAL_NOSET_CUDA_VISIBLE_DEVICES: '1'
RAY_EXPERIMENTAL_NOSET_HABANA_VISIBLE_MODULES: '1'
RAY_EXPERIMENTAL_NOSET_HIP_VISIBLE_DEVICES: '1'
RAY_EXPERIMENTAL_NOSET_NEURON_RT_VISIBLE_CORES: '1'
RAY_EXPERIMENTAL_NOSET_ONEAPI_DEVICE_SELECTOR: '1'
RAY_EXPERIMENTAL_NOSET_TPU_VISIBLE_CHIPS: '1'
SGLANG_BATCH_INVARIANT_OPS_ENABLE_MM_FALLBACK_VARIANT: 'true'
SGLANG_DG_CACHE_DIR_PER_PROCESS: '1'
SGLANG_ENABLE_HEALTH_ENDPOINT_GENERATION: 'false'
SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_IDLE: 'false'
SGLANG_ENABLE_TP_MEMORY_INBALANCE_CHECK: 'false'
SGLANG_EXPOSE_OWN_ENV_VARS: '1'
SGLANG_JIT_DEEPGEMM_PRECOMPILE: 'false'
SGLANG_MEMORY_SAVER_CUDA_GRAPH: 'true'
SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2: '1'
meta:
gpu_ids: 0,1,2,3,4,5,6,7
name: inference-engine-all-0-1
objectName: myrun-miles-run-inference-engine-all-0-1
poolId: inference-engine-all-0-1
ports:
- name: primary
port: 8000
- name: dist-init
port: 9000
- name: nccl
port: 10000
- name: engine-info-boo
port: 12000
- name: gate
port: 13000
replicas: 2
resources:
limits:
nvidia.com/gpu: 8
objectNames:
colocatePairing: myrun-miles-run-colocate-pairing
mooncakeMaster: myrun-miles-run-mooncake-master
orchestrator: myrun-miles-run-orchestrator
platformRead: myrun-miles-run-platform-read
platformReadDelete: myrun-miles-run-platform-read-delete
uninstall: myrun-miles-run-uninstall
uninstallManifest: myrun-miles-run-uninstall-manifest
orchestrator:
command:
- python
- scripts/run_qwen3_4b.py
- train
- --cluster-backend
- kubernetes
stateFile: /cluster-storage/myteam/miles_data/miles-runs/260101-000000-000/state/orchestrator-260101-000000-000001.state
staticWorkers:
- command:
- <PYTHON>
- -m
- miles.utils.workers.serving.serve
- --specs
- miles.ray.specs.entrypoint.compute_specs_from_argv
- --pool-id
- rollout-executor
- --
- --cluster-backend
- kubernetes
- --rollout-num-gpus
- '32'
name: rollout-executor
objectName: myrun-miles-run-rollout-executor
poolId: rollout-executor
ports:
- name: rpc
port: 8000
replicas: 1
- command:
- <PYTHON>
- -m
- miles.utils.workers.serving.serve
- --specs
- miles.ray.specs.entrypoint.compute_specs_from_argv
- --pool-id
- inference-controller
- --
- --cluster-backend
- kubernetes
- --rollout-num-gpus
- '32'
name: inference-controller
objectName: myrun-miles-run-inference-controller
poolId: inference-controller
ports:
- name: rpc
port: 8000
replicas: 1
serviceAccountName: myrun-miles-run-platform-read
- command:
- <PYTHON>
- -m
- sglang_router.launch_router
- --policy
- manual
- --assignment-mode
- min_load
- --pd-disaggregation
- --log-level
- warn
- --prometheus-port
- '9000'
- --request-timeout-secs
- '14400'
- --port
- '8000'
name: inference-router-0
objectName: myrun-miles-run-inference-router-0
poolId: inference-router-0
ports:
- name: primary
port: 8000
- name: prometheus
port: 9000
replicas: 1
- command:
- <PYTHON>
- -m
- miles.rollout.session.server
- --config-json
- '{"host":"0.0.0.0","port":8000,"instance_id":"0123456789abcdef-0","backend_url":"http://myrun-miles-run-inference-router-0-0.myrun-miles-run-inference-router-0:8000","timeout":null,"hf_checkpoint":"<FIXTURE_DIR>/typical-model","chat_template_path":null,"tito_model":"default","apply_chat_template_kwargs":{},"use_rollout_routing_replay":false,"use_rollout_indexer_replay":false,"use_sampling_support_replay":false,"sglang_speculative_algorithm":null,"num_layers":36,"moe_router_topk":2,"save_debug_trajectory_data":null,"lora_rank":0,"lora_adapter_path":null,"lora_train_only":false,"use_session_server":true,"session_message_matcher":"strict","pause_generation_mode":"retract","session_sample_picker_path":"miles.rollout.session.v2.picker_hub.drop_same_prompt_retries","session_sample_postprocessor_path":"miles.rollout.session.v2.postprocessor_hub.default_postprocess"}'
name: session-server
objectName: myrun-miles-run-session-server
poolId: session-server
ports:
- name: primary
port: 8000
replicas: 32
- command:
- <PYTHON>
- -m
- miles.utils.workers.serving.serve
- --specs
- miles.ray.specs.entrypoint.compute_specs_from_argv
- --pool-id
- trainer-controller-actor
- --
- --cluster-backend
- kubernetes
- --rollout-num-gpus
- '32'
name: trainer-controller-actor
objectName: myrun-miles-run-trainer-controller-actor
poolId: trainer-controller-actor
ports:
- name: rpc
port: 8000
replicas: 1
serviceAccountName: myrun-miles-run-platform-read-delete
trainerEngines:
- command:
- bash
- -c
- mkdir -p /scratch/Qwen3-4B && rsync -a --info=progress2 /cluster-storage/models/Qwen3-4B/ /scratch/Qwen3-4B && exec <PYTHON> -m miles.utils.workers.process_supervisor --num-subprocesses 8 -- <PYTHON> -m miles.utils.workers.serving.serve --specs miles.ray.specs.entrypoint.compute_specs_from_argv --pool-id trainer-engine-actor -- --cluster-backend kubernetes --rollout-num-gpus 32
meta:
gpu_ids: 0,1,2,3,4,5,6,7
name: trainer-engine-actor
objectName: myrun-miles-run-trainer-engine-actor
poolId: trainer-engine-actor
ports:
- name: master
port: 9000
- name: rpc
port: 8000
replicas: 1
resources:
limits:
nvidia.com/gpu: 8
size: 4
Reference in New Issue View Git Blame Copy Permalink